xen/privcmd: improve performance of MMAPBATCH_V2

Make the IOCTL_PRIVCMD_MMAPBATCH_V2 (and older V1 version) map
multiple frames at a time rather than one at a time, despite the pages
being non-consecutive GFNs.

xen_remap_foreign_mfn_array() is added which maps an array of GFNs
(instead of a consecutive range of GFNs).

Since per-frame errors are returned in an array, privcmd must set the
MMAPBATCH_V1 error bits as part of the "report errors" phase, after
all the frames are mapped.

Migrate times are significantly improved (when using a PV toolstack
domain).  For example, for an idle 12 GiB PV guest:

        Before     After
  real  0m38.179s  0m26.868s
  user  0m15.096s  0m13.652s
  sys   0m28.988s  0m18.732s

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>

2 files changed, 110 insertions, 53 deletions

diff --git a/drivers/xen/privcmd.c b/drivers/xen/privcmd.c
index 59ac71c4a043..5a296161d843 100644
--- a/drivers/xen/privcmd.c
+++ b/drivers/xen/privcmd.c
@@ -159,6 +159,40 @@ static int traverse_pages(unsigned nelem, size_t size,
 	return ret;
 }
 
+/*
+ * Similar to traverse_pages, but use each page as a "block" of
+ * data to be processed as one unit.
+ */
+static int traverse_pages_block(unsigned nelem, size_t size,
+				struct list_head *pos,
+				int (*fn)(void *data, int nr, void *state),
+				void *state)
+{
+	void *pagedata;
+	unsigned pageidx;
+	int ret = 0;
+
+	BUG_ON(size > PAGE_SIZE);
+
+	pageidx = PAGE_SIZE;
+
+	while (nelem) {
+		int nr = (PAGE_SIZE/size);
+		struct page *page;
+		if (nr > nelem)
+			nr = nelem;
+		pos = pos->next;
+		page = list_entry(pos, struct page, lru);
+		pagedata = page_address(page);
+		ret = (*fn)(pagedata, nr, state);
+		if (ret)
+			break;
+		nelem -= nr;
+	}
+
+	return ret;
+}
+
 struct mmap_mfn_state {
 	unsigned long va;
 	struct vm_area_struct *vma;
@@ -274,39 +308,25 @@ struct mmap_batch_state {
 /* auto translated dom0 note: if domU being created is PV, then mfn is
  * mfn(addr on bus). If it's auto xlated, then mfn is pfn (input to HAP).
  */
-static int mmap_batch_fn(void *data, void *state)
+static int mmap_batch_fn(void *data, int nr, void *state)
 {
 	xen_pfn_t *mfnp = data;
 	struct mmap_batch_state *st = state;
 	struct vm_area_struct *vma = st->vma;
 	struct page **pages = vma->vm_private_data;
-	struct page *cur_page = NULL;
+	struct page **cur_pages = NULL;
 	int ret;
 
 	if (xen_feature(XENFEAT_auto_translated_physmap))
-		cur_page = pages[st->index++];
+		cur_pages = &pages[st->index];
 
-	ret = xen_remap_domain_mfn_range(st->vma, st->va & PAGE_MASK, *mfnp, 1,
-					 st->vma->vm_page_prot, st->domain,
-					 &cur_page);
+	BUG_ON(nr < 0);
+	ret = xen_remap_domain_mfn_array(st->vma, st->va & PAGE_MASK, mfnp, nr,
+					 (int *)mfnp, st->vma->vm_page_prot,
+					 st->domain, cur_pages);
 
-	/* Store error code for second pass. */
-	if (st->version == 1) {
-		if (ret < 0) {
-			/*
-			 * V1 encodes the error codes in the 32bit top nibble of the
-			 * mfn (with its known limitations vis-a-vis 64 bit callers).
-			 */
-			*mfnp |= (ret == -ENOENT) ?
-						PRIVCMD_MMAPBATCH_PAGED_ERROR :
-						PRIVCMD_MMAPBATCH_MFN_ERROR;
-		}
-	} else { /* st->version == 2 */
-		*((int *) mfnp) = ret;
-	}
-
-	/* And see if it affects the global_error. */
-	if (ret < 0) {
+	/* Adjust the global_error? */
+	if (ret != nr) {
 		if (ret == -ENOENT)
 			st->global_error = -ENOENT;
 		else {
@@ -315,23 +335,35 @@ static int mmap_batch_fn(void *data, void *state)
 				st->global_error = 1;
 		}
 	}
-	st->va += PAGE_SIZE;
+	st->va += PAGE_SIZE * nr;
+	st->index += nr;
 
 	return 0;
 }
 
-static int mmap_return_errors(void *data, void *state)
+static int mmap_return_error(int err, struct mmap_batch_state *st)
 {
-	struct mmap_batch_state *st = state;
+	int ret;
 
 	if (st->version == 1) {
-		xen_pfn_t mfnp = *((xen_pfn_t *) data);
-		if (mfnp & PRIVCMD_MMAPBATCH_MFN_ERROR)
-			return __put_user(mfnp, st->user_mfn++);
-		else
+		if (err) {
+			xen_pfn_t mfn;
+
+			ret = get_user(mfn, st->user_mfn);
+			if (ret < 0)
+				return ret;
+			/*
+			 * V1 encodes the error codes in the 32bit top
+			 * nibble of the mfn (with its known
+			 * limitations vis-a-vis 64 bit callers).
+			 */
+			mfn |= (err == -ENOENT) ?
+				PRIVCMD_MMAPBATCH_PAGED_ERROR :
+				PRIVCMD_MMAPBATCH_MFN_ERROR;
+			return __put_user(mfn, st->user_mfn++);
+		} else
 			st->user_mfn++;
 	} else { /* st->version == 2 */
-		int err = *((int *) data);
 		if (err)
 			return __put_user(err, st->user_err++);
 		else
@@ -341,6 +373,21 @@ static int mmap_return_errors(void *data, void *state)
 	return 0;
 }
 
+static int mmap_return_errors(void *data, int nr, void *state)
+{
+	struct mmap_batch_state *st = state;
+	int *errs = data;
+	int i;
+	int ret;
+
+	for (i = 0; i < nr; i++) {
+		ret = mmap_return_error(errs[i], st);
+		if (ret < 0)
+			return ret;
+	}
+	return 0;
+}
+
 /* Allocate pfns that are then mapped with gmfns from foreign domid. Update
  * the vma with the page info to use later.
  * Returns: 0 if success, otherwise -errno
@@ -472,8 +519,8 @@ static long privcmd_ioctl_mmap_batch(void __user *udata, int version)
 	state.version       = version;
 
 	/* mmap_batch_fn guarantees ret == 0 */
-	BUG_ON(traverse_pages(m.num, sizeof(xen_pfn_t),
-			     &pagelist, mmap_batch_fn, &state));
+	BUG_ON(traverse_pages_block(m.num, sizeof(xen_pfn_t),
+				    &pagelist, mmap_batch_fn, &state));
 
 	up_write(&mm->mmap_sem);
 
@@ -481,8 +528,8 @@ static long privcmd_ioctl_mmap_batch(void __user *udata, int version)
 		/* Write back errors in second pass. */
 		state.user_mfn = (xen_pfn_t *)m.arr;
 		state.user_err = m.err;
-		ret = traverse_pages(m.num, sizeof(xen_pfn_t),
-							 &pagelist, mmap_return_errors, &state);
+		ret = traverse_pages_block(m.num, sizeof(xen_pfn_t),
+					   &pagelist, mmap_return_errors, &state);
 	} else
 		ret = 0;
 
diff --git a/drivers/xen/xlate_mmu.c b/drivers/xen/xlate_mmu.c
index 7724d90fc697..58a5389aec89 100644
--- a/drivers/xen/xlate_mmu.c
+++ b/drivers/xen/xlate_mmu.c
@@ -62,13 +62,15 @@ static int map_foreign_page(unsigned long lpfn, unsigned long fgmfn,
 }
 
 struct remap_data {
-	xen_pfn_t fgmfn; /* foreign domain's gmfn */
+	xen_pfn_t *fgmfn; /* foreign domain's gmfn */
 	pgprot_t prot;
 	domid_t  domid;
 	struct vm_area_struct *vma;
 	int index;
 	struct page **pages;
 	struct xen_remap_mfn_info *info;
+	int *err_ptr;
+	int mapped;
 };
 
 static int remap_pte_fn(pte_t *ptep, pgtable_t token, unsigned long addr,
@@ -80,38 +82,46 @@ static int remap_pte_fn(pte_t *ptep, pgtable_t token, unsigned long addr,
 	pte_t pte = pte_mkspecial(pfn_pte(pfn, info->prot));
 	int rc;
 
-	rc = map_foreign_page(pfn, info->fgmfn, info->domid);
-	if (rc < 0)
-		return rc;
-	set_pte_at(info->vma->vm_mm, addr, ptep, pte);
+	rc = map_foreign_page(pfn, *info->fgmfn, info->domid);
+	*info->err_ptr++ = rc;
+	if (!rc) {
+		set_pte_at(info->vma->vm_mm, addr, ptep, pte);
+		info->mapped++;
+	}
+	info->fgmfn++;
 
 	return 0;
 }
 
-int xen_xlate_remap_gfn_range(struct vm_area_struct *vma,
+int xen_xlate_remap_gfn_array(struct vm_area_struct *vma,
 			      unsigned long addr,
-			      xen_pfn_t gfn, int nr,
-			      pgprot_t prot, unsigned domid,
+			      xen_pfn_t *mfn, int nr,
+			      int *err_ptr, pgprot_t prot,
+			      unsigned domid,
 			      struct page **pages)
 {
 	int err;
 	struct remap_data data;
+	unsigned long range = nr << PAGE_SHIFT;
 
-	/* TBD: Batching, current sole caller only does page at a time */
-	if (nr > 1)
-		return -EINVAL;
+	/* Kept here for the purpose of making sure code doesn't break
+	   x86 PVOPS */
+	BUG_ON(!((vma->vm_flags & (VM_PFNMAP | VM_IO)) == (VM_PFNMAP | VM_IO)));
 
-	data.fgmfn = gfn;
-	data.prot = prot;
+	data.fgmfn = mfn;
+	data.prot  = prot;
 	data.domid = domid;
-	data.vma = vma;
-	data.index = 0;
+	data.vma   = vma;
 	data.pages = pages;
-	err = apply_to_page_range(vma->vm_mm, addr, nr << PAGE_SHIFT,
+	data.index = 0;
+	data.err_ptr = err_ptr;
+	data.mapped = 0;
+
+	err = apply_to_page_range(vma->vm_mm, addr, range,
 				  remap_pte_fn, &data);
-	return err;
+	return err < 0 ? err : data.mapped;
 }
-EXPORT_SYMBOL_GPL(xen_xlate_remap_gfn_range);
+EXPORT_SYMBOL_GPL(xen_xlate_remap_gfn_array);
 
 int xen_xlate_unmap_gfn_range(struct vm_area_struct *vma,
 			      int nr, struct page **pages)